API vs Self-Hosted GPU Break-Even Calculator
Find the exact monthly token crossover where renting dedicated H100, H200, A100, or L40S GPUs with vLLM becomes cheaper than paying per-token API rates—including KV-cache, HA redundancy, and DevOps overhead.
Ready-to-Run vLLM Cluster Command
Open-Weight Break-Even Directory (FP8, 24/7 Single Cluster + $400 Ops)
| Model | Params (Active) | FP8 VRAM | Min Cloud GPU Cluster | Self-Host / Mo | API Blended | Break-Even Volume |
|---|---|---|---|---|---|---|
| Google: Gemma 4 26B A4B (free) | 26B (4B) | 50.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | Free | Free API ($0) |
| Google: Gemma 4 31B (free) | 31B | 58.7 GB | 1x NVIDIA A100 80GB SXM | $1,488 | Free | Free API ($0) |
| Meta: Llama 3.1 8B Instruct | 8B | 15.6 GB | 1x NVIDIA RTX 4090 | $721.00 | $0.025 | 28,840M tok/mo |
| Mistral: Mistral Nemo | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.029 | 50,785M tok/mo |
| Qwen: Qwen3.7 Flash | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.055 | 27,054.5M tok/mo |
| Google: Gemma 3 4B | 4B | 7.8 GB | 1x NVIDIA RTX 4090 | $721.00 | $0.063 | 11,536M tok/mo |
| DeepSeek: DeepSeek V4 Flash Latest | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.066 | 22,545.5M tok/mo |
| Google: Gemma 3 12B | 12B | 23.4 GB | 2x NVIDIA RTX 4090 | $1,042 | $0.075 | 13,893.3M tok/mo |
| DeepSeek: DeepSeek V4.1 Flash | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.087 | 17,083.8M tok/mo |
| Microsoft: Phi 4 | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.087 | 17,005.7M tok/mo |
| Google: Gemma 4 26B A4B | 26B (4B) | 50.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.107 | 13,919.6M tok/mo |
| DeepSeek: DeepSeek V4 Flash 0731 | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.108 | 13,752.3M tok/mo |
| DeepSeek: DeepSeek V4 Flash 0423 | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.113 | 13,226.7M tok/mo |
| Qwen: Qwen3.5-Flash | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.114 | 13,075.6M tok/mo |
| Meta: Llama 3.2 3B Instruct | 3B | 5.9 GB | 1x NVIDIA RTX 4090 | $721.00 | $0.12 | 6,008.3M tok/mo |
| Qwen: Qwen3 Coder 30B A3B Instruct | 30B (3B) | 57.6 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.122 | 12,146.9M tok/mo |
| Qwen: Qwen3 32B | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.13 | 11,446.2M tok/mo |
| Meta: Llama 4 Scout | 109B (17B) | 164.8 GB | 8x NVIDIA RTX 4090 | $2,970 | $0.15 | 19,800M tok/mo |
| Qwen: Qwen3 14B | 14B | 27.3 GB | 2x NVIDIA RTX 4090 | $1,042 | $0.15 | 6,946.7M tok/mo |
| Qwen: Qwen3 30B A3B Instruct 2507 | 30B (3B) | 57.6 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.15 | 9,920M tok/mo |
| DeepSeek: DeepSeek Flash Latest | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.152 | 9,789.5M tok/mo |
| Google: Gemma 3 27B | 27B | 52.7 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.172 | 8,626.1M tok/mo |
| Meta: Llama Guard 4 12B | 12B | 23.4 GB | 2x NVIDIA RTX 4090 | $1,042 | $0.18 | 5,788.9M tok/mo |
| Qwen: Qwen3.5-9B | 9B | 17.6 GB | 1x NVIDIA RTX 4090 | $721.00 | $0.19 | 3,794.7M tok/mo |
| Qwen: Qwen3 30B A3B | 30B (3B) | 57.6 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.215 | 6,920.9M tok/mo |
| Qwen: Qwen3.8 Flash | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.23 | 6,469.6M tok/mo |
| Qwen: Qwen3.8 Omni Flash | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.23 | 6,469.6M tok/mo |
| Google: Gemma 4 31B | 31B | 58.7 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.25 | 5,952M tok/mo |
| Cohere: Command R (08-2024) | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.263 | 5,668.6M tok/mo |
| DeepSeek: DeepSeek V4.1 Flash 1017 | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.263 | 5,668.6M tok/mo |
| Qwen: Qwen3 VL 30B A3B Instruct | 30B (3B) | 57.6 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.263 | 5,668.6M tok/mo |
| DeepSeek: DeepSeek V3.2 | 685B (37B) | 1035.7 GB | 16x NVIDIA A100 80GB SXM | $17,803 | $0.29 | 61,389.7M tok/mo |
| Qwen: Qwen3 Coder Next | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.29 | 5,131M tok/mo |
| Qwen: Qwen3.6 35B A3B | 35B (3B) | 63.2 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.30 | 4,960M tok/mo |
| DeepSeek: DeepSeek V3.2 Exp | 685B (37B) | 1035.7 GB | 16x NVIDIA A100 80GB SXM | $17,803 | $0.305 | 58,370.5M tok/mo |
| Qwen: Qwen3 235B A22B Instruct 2507 | 235B (22B) | 355.3 GB | 8x NVIDIA A100 80GB SXM | $9,102 | $0.305 | 29,842.6M tok/mo |
| DeepSeek: DeepSeek V4 Flash Vision Exp | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.323 | 4,601.1M tok/mo |
| Qwen: Qwen3 Next 80B A3B Instruct | 80B (3B) | 121 GB | 2x NVIDIA A100 80GB SXM | $2,575 | $0.35 | 7,357.1M tok/mo |
| Qwen: Qwen3.5-35B-A3B | 35B (3B) | 63.2 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.355 | 4,191.5M tok/mo |
| Qwen2.5 72B Instruct | 72B | 108.9 GB | 2x NVIDIA A100 80GB SXM | $2,575 | $0.37 | 6,959.5M tok/mo |
| Qwen: Qwen3 VL 8B Instruct | 8B | 15.6 GB | 1x NVIDIA RTX 4090 | $721.00 | $0.375 | 1,922.7M tok/mo |
| Qwen: Qwen-Plus | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.39 | 3,815.4M tok/mo |
| Qwen: Qwen3 Coder Flash | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.39 | 3,815.4M tok/mo |
| Qwen: Qwen3.6 Flash | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.422 | 3,526.9M tok/mo |
| DeepSeek: DeepSeek V3.1 | 685B (37B) | 1035.7 GB | 16x NVIDIA A100 80GB SXM | $17,803 | $0.425 | 41,889.4M tok/mo |
| DeepSeek: DeepSeek V3 0324 | 685B (37B) | 1035.7 GB | 16x NVIDIA A100 80GB SXM | $17,803 | $0.438 | 40,692.6M tok/mo |
| DeepSeek: DeepSeek V3 | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.45 | 3,305.2M tok/mo |
| DeepSeek: DeepSeek V3.1 Terminus | 685B (37B) | 1035.7 GB | 16x NVIDIA A100 80GB SXM | $17,803 | $0.453 | 39,343.6M tok/mo |
| Qwen: Qwen3 Coder 480B A35B | 480B (35B) | 725.8 GB | 16x NVIDIA A100 80GB SXM | $17,803 | $0.475 | 37,480M tok/mo |
| Meta: Llama 4 Maverick | 400B (17B) | 604.8 GB | 16x NVIDIA L40S 48GB | $10,795 | $0.512 | 21,063.4M tok/mo |
| Qwen: Qwen3.8 27B | 27B | 52.7 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.535 | 2,781.3M tok/mo |
| Qwen: Qwen3.5-27B | 27B | 52.7 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.536 | 2,774.6M tok/mo |
| Qwen: Qwen3.7 Plus | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.56 | 2,657.1M tok/mo |
| Qwen: Qwen3.5 Plus 2026-02-15 | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.585 | 2,543.6M tok/mo |
| Qwen: Qwen3 VL 235B A22B Instruct | 235B (22B) | 355.3 GB | 8x NVIDIA A100 80GB SXM | $9,102 | $0.632 | 14,390.5M tok/mo |
| Google: Gemma 2 27B | 27B | 41.5 GB | 2x NVIDIA RTX 4090 | $1,042 | $0.65 | 1,603.1M tok/mo |
| Qwen: Qwen3.5 Plus 2026-04-20 | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.675 | 2,204.4M tok/mo |
| Meta: Llama 3.3 70B Instruct | 70B | 105.8 GB | 2x NVIDIA A100 80GB SXM | $2,575 | $0.71 | 3,626.8M tok/mo |
| Qwen: Qwen3.5-122B-A10B | 122B (10B) | 184.5 GB | 4x NVIDIA A100 80GB SXM | $4,751 | $0.715 | 6,644.8M tok/mo |
| Meta: Llama 3.1 70B Instruct | 70B | 105.8 GB | 2x NVIDIA A100 80GB SXM | $2,575 | $0.72 | 3,576.4M tok/mo |
| Qwen: Qwen3.6 Plus | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.731 | 2,034.7M tok/mo |
| Qwen2.5 Coder 32B Instruct | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $0.745 | 1,997.3M tok/mo |
| Qwen: Qwen2.5 VL 72B Instruct | 72B | 108.9 GB | 2x NVIDIA A100 80GB SXM | $2,575 | $0.85 | 3,029.4M tok/mo |
| DeepSeek: R1 0528 | 685B (37B) | 1035.7 GB | 16x NVIDIA A100 80GB SXM | $17,803 | $1.00 | 17,806.6M tok/mo |
| AionLabs: Aion-RP 1.0 (8B) | 8B | 15.6 GB | 1x NVIDIA RTX 4090 | $721.00 | $1.00 | 721M tok/mo |
| Qwen: Qwen3.6 27B | 27B | 52.7 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $1.04 | 1,430.8M tok/mo |
| DeepSeek: DeepSeek V4 Pro 0423 | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $1.05 | 1,417.1M tok/mo |
| Qwen: Qwen3.5 397B A17B | 397B (17B) | 600.3 GB | 16x NVIDIA L40S 48GB | $10,795 | $1.09 | 9,926.4M tok/mo |
| Qwen: Qwen3 235B A22B Thinking 2507 | 235B (22B) | 355.3 GB | 8x NVIDIA A100 80GB SXM | $9,102 | $1.21 | 7,506.8M tok/mo |
| Nous: Hermes 4 405B | 405B | 612.4 GB | 16x NVIDIA L40S 48GB | $10,795 | $1.50 | 7,196.7M tok/mo |
| DeepSeek: DeepSeek V4 Pro 0813 | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $1.98 | 751.5M tok/mo |
| DeepSeek: DeepSeek Pro Latest | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $2.07 | 717.1M tok/mo |
| Qwen: Qwen4 Max | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $2.14 | 696.1M tok/mo |
| Qwen: Qwen3.7 Max | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $2.21 | 672.5M tok/mo |
| Mistral: Mixtral 8x22B Instruct | 141B (39B) | 213.2 GB | 4x NVIDIA A100 80GB SXM | $4,751 | $3.00 | 1,583.7M tok/mo |
| Qwen: Qwen3.8 2.4T A95B | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $3.00 | 496M tok/mo |
| Qwen: Qwen3.8 Max (0902) | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $3.00 | 496M tok/mo |
| Cohere: Command R+ (08-2024) | 104B | 157.2 GB | 8x NVIDIA RTX 4090 | $2,970 | $4.38 | 678.9M tok/mo |
| Qwen: Qwen3.8 Max Prime | 32B | 59.8 GB | 1x NVIDIA A100 80GB SXM | $1,488 | $6.00 | 248M tok/mo |
Launch Dedicated GPUs or Route via Serverless API
RunPod
On-demand H100/A100 GPU pods and serverless vLLM endpoints for private open-weight hosting.
DigitalOcean GPU Droplets
Simple cloud compute and managed inference droplets with predictable monthly pricing.
Together AI
High-throughput serverless inference & fine-tuning for Llama, DeepSeek, and Qwen.
OpenRouter
Unified API gateway for 300+ LLMs with automatic fallbacks and unified billing.
Some outbound infrastructure and API gateway links on KostoZero are partner referrals that help keep this observatory free and independent at no extra cost to you.
Frequently Asked Questions
When is self-hosting an LLM on GPUs cheaper than paying per-token APIs?
Self-hosting wins once your monthly token spend exceeds fixed GPU rental + DevOps maintenance + HA redundancy overhead. Because serverless open-weight APIs now cost under $0.30/1M tokens, most teams need 150M–1,500M+ steady tokens/month (or strict privacy rules) to beat managed APIs.
Why does KostoZero separate Total vs Active parameters for MoE models?
Mixture-of-Experts models like DeepSeek-V3 (685B total / 37B active) or Llama 4 Maverick (400B total / 17B active) must fit ALL weights into GPU VRAM, while inference memory bandwidth and decode speed scale with ACTIVE parameters.
Should I choose FP8 or INT4 quantization for production vLLM serving?
FP8 on Hopper/Ada/Blackwell GPUs (H100, H200, L40S, B200) cuts VRAM nearly in half with negligible accuracy loss. INT4/AWQ fits 70B+ models on fewer GPUs when minimizing idle hardware cost matters most.
Get Instant Price-Drop (>10%), New AI Bot & SaaS Cliff Alerts
Every 6 hours our autonomous sync checks 40+ LLMs, AI crawlers, and SaaS pricing cliffs—alerting you when prices drop >10% or new scrapers appear.